US11436795B2ActiveUtilityPatentIndex 48
Learning a neural network for inference of editable feature trees
Est. expiryDec 29, 2038(~12.5 yrs left)· nominal 20-yr term from priority
G06N 3/045G06N 3/044G06N 3/09G06N 3/092G06N 3/0442G06N 3/0455G06N 3/0464G06N 3/006G06N 5/041G06N 3/084G06T 17/005G06F 3/0484G06T 17/10G06N 3/08G06N 3/0445
48
PatentIndex Score
0
Cited by
12
References
20
Claims
Abstract
The disclosure notably relates to a computer-implemented method for learning a neural network configured for inference, from a discrete geometrical representation of a 3D shape, of an editable feature tree representing the 3D shape. The editable feature tree includes a tree arrangement of geometrical operations applied to leaf geometrical shapes. The method includes obtaining a dataset including discrete geometrical representations each of a respective 3D shape, and obtaining a candidate set of leaf geometrical shapes. The method also includes learning the neural network based on the dataset and on the candidate set. The candidate set includes at least one continuous subset of leaf geometrical shapes. The method forms an improved solution for digitization.
Claims
exact text as granted — not AI-modifiedThe invention claimed is:
1. A computer-implemented method for learning a neural network configured for inference, from a discrete geometrical representation of a 3D shape, of an editable feature tree representing the 3D shape, the editable feature tree comprising a tree arrangement of geometrical operations applied to leaf geometrical shapes, the method comprising:
obtaining a dataset including discrete geometrical representations each of a respective 3D shape;
obtaining a candidate set of leaf geometrical shapes, each leaf geometrical shape having a respective 3D shape, the candidate set comprising at least one continuous subset of leaf geometrical shapes, each first candidate leaf geometrical shape of the subset being obtainable by a continuous deformation of second candidate leaf geometrical shape of the subset, the continuous deformation continuously modifying the respective 3D shape of the first candidate leaf geometrical shape into the respective 3D shape of the second candidate leaf geometrical shape, such that each intermediate state of the continuous deformation is itself a candidate leaf geometrical shape of the subset, the intermediate state having a respective 3D shape which is intermediate between the respective 3D shape of the first candidate leaf geometrical shape and the respective 3D shape of the second candidate leaf geometrical shape; and
learning the neural network based on the dataset and on the candidate set such that each leaf geometrical shape of the continuous subset of leaf geometrical shapes is available for being potentially inferred during inference by the learnt neural network,
wherein the candidate set includes a set product between:
a discrete set of primitive shape types, and
for each primitive shape type, a respective discrete set of one or more parameter domains each of a respective continuous parameter, each parameter domain having respective parameters values of the respective continuous parameter,
each primitive shape type forming, with a respective parameter value for each of the respective discrete set of one or more parameter domains, a respective element of the candidate set, and
wherein for each of one or more primitive shape types, the one or more respective continuous parameters include one or more dimensional parameters, the dimensional parameters being parameters that define a size of characteristic dimensions of a shape of a given primitive shape type.
2. The method of claim 1 , wherein, for each of one or more primitive shape types, the one or more respective continuous parameters include one or more dimensional parameters and/or one or more positioning parameters.
3. The method of claim 1 , wherein the discrete set of primitive shape types includes a cuboid type, a sphere type, one or more cylinder types, and/or one or more prism types.
4. The method of claim 1 , wherein the neural network includes recurrent neural network cells, each RNN cell outputting at a respective time step, respective first data for inference of a respective primitive shape type and respective second data for inference of a respective parameter value for each of the respective discrete set of one or more parameter domains.
5. The method of claim 4 , wherein the respective first data include a respective discrete distribution of probabilities each attributed to a respective one of the discrete set of primitive shape types, and/or the respective second data comprise a respective parameter value for each of the respective discrete set of one or more parameter domains.
6. The method of claim 5 , wherein the dataset further includes, for each of one or more discrete geometrical representations, a respective editable feature tree comprising a tree arrangement of geometrical operations applied to leaf geometrical shapes and representing the 3D shape respective to the discrete geometrical representation, each geometrical shape being formed by a respective primitive shape type with a respective parameter value for each of the respective discrete set of one or more parameter domains, the learning of the neural network comprising a supervised training which includes minimizing a loss L 1 which penalizes, for the time step respective to each leaf geometrical shape of each discrete geometrical representation:
a lowness of the probability of the respective first data attributed to the respective primitive shape type of the leaf geometrical shape, and/or
a disparity between the one or more respective parameter values of the leaf geometrical shape and the one or more respective parameter values of the respective second data.
7. The method of claim 6 , wherein the loss L 1 penalizes:
the lowness with a term of the type
-
∑
n
=
1
N
1
T
^
n
∑
t
=
1
T
^
n
log
[
]
;
and/or
the disparity with a term of the type
∑
n
=
1
N
1
T
^
n
∑
t
=
1
T
^
n
incr
[
d
(
,
)
]
;
where:
n designates said discrete geometrical representation,
t designates said time step,
designates said respective primitive shape type,
designates the probability of the respective first data attributed to ,
designates said one or more respective parameter values of said leaf geometrical shape,
designates said one or more respective parameter values of the respective second data,
d(⋅,⋅) is a distance function, and
incr[⋅] is an increasing and/or positive function.
8. The method of claim 5 , wherein the respective first data include a respective discrete distribution of probabilities each attributed to a respective one of the discrete set of primitive shape types, the learning of the neural network including an unsupervised training which includes minimizing a loss L 2 , the minimizing including exploring candidate respective discrete distributions of probabilities, the loss L 2 penalizing, for a discrete geometrical representation and for a candidate respective discrete distribution of probabilities, a disparity with a discrete geometrical representation derived from a respective editable feature tree, the respective editable feature tree being inferable based on the explored candidate respective discrete distribution of probabilities, the minimizing including a backpropagation of a gradient of the loss L 2 relative at least to a variable representing the candidate respective discrete distributions of the respective first data, the gradient being obtained with a reinforcement algorithm.
9. The method of claim 8 , wherein the gradient is of the type
∂
L
2
∂
p
y
=
L
2
-
β
p
y
1
t
≤
T
n
1
y
=
a
t
n
,
where:
n designates said discrete geometrical representation,
T n designates a depth of said respective editable feature tree,
t designates a respective time step,
y designates a primitive shape type,
p y designates said variable representing the candidate respective discrete distributions of the respective first data at t,
a t n designates a primitive shape type of said respective editable feature tree inferable at t, and
β is a baseline value.
10. The method of claim 4 , wherein each RNN cell takes as input a result of a current predicted sequence, the learning of the neural network including a supervised training which includes minimizing a loss that involves inferred parameter values.
11. A computer-implemented method of applying a data structure representing a neural network learnable according to a computer-implemented process for learning a neural network configured for inference, from a discrete geometrical representation of a 3D shape, of an editable feature tree representing the 3D shape, the editable feature tree comprising a tree arrangement of geometrical operations applied to leaf geometrical shapes, the method comprising:
obtaining a dataset including discrete geometrical representations each of a respective 3D shape;
obtaining a candidate set of leaf geometrical shapes, each leaf geometrical shape having a respective 3D shape, the candidate set comprising at least one continuous subset of leaf geometrical shapes, each first candidate leaf geometrical shape of the subset being obtainable by a continuous deformation of second candidate leaf geometrical shape of the subset, the continuous deformation continuously modifying the respective 3D shape of the first candidate leaf geometrical shape into the respective 3D shape of the second candidate leaf geometrical shape, such that each intermediate state of the continuous deformation is itself a candidate leaf geometrical shape of the subset, the intermediate state having a respective 3D shape which is intermediate between the respective 3D shape of the first candidate leaf geometrical shape and the respective 3D shape of the second candidate leaf geometrical shape;
learning the neural network based on the dataset and on the candidate set such that each leaf geometrical shape of the continuous subset of leaf geometrical shapes is available for being potentially inferred during inference by the learnt neural network;
obtaining a discrete geometrical representation of a 3D shape;
applying the neural network to the discrete geometrical representation of a 3D shape; and
inferring an editable feature tree representing the 3D shape based on a result of the applying,
wherein the candidate set includes a set product between:
a discrete set of primitive shape types, and
for each primitive shape type, a respective discrete set of one or more parameter domains each of a respective continuous parameter, each parameter domain having respective parameters values of the respective continuous parameter,
each primitive shape type forming, with a respective parameter value for each of the respective discrete set of one or more parameter domains, a respective element of the candidate set, and
wherein for each of one or more primitive shape types, the one or more respective continuous parameters include one or more dimensional parameters, the dimensional parameters being parameters that define a size of characteristic dimensions of a shape of a given primitive shape type.
12. A device comprising:
a processor; and
a non-transitory data storage medium having recorded thereon:
a data structure representing a neural network learnable according to a computer-implemented process for learning a neural network configured for inference, from a discrete geometrical representation of a 3D shape, of an editable feature tree representing the 3D shape, the editable feature tree comprising a tree arrangement of geometrical operations applied to leaf geometrical shapes,
a computer program comprising instruction for learning the neural network configured for inference, from the discrete geometrical representation of the 3D shape, of the editable feature tree representing the 3D shape, the editable feature tree comprising the tree arrangement of geometrical operations applied to leaf geometrical shapes,
wherein the computer program when executed by the processor causes the processor to be configured to
obtain a dataset including discrete geometrical representations each of a respective 3D shape,
obtaining a candidate set of leaf geometrical shapes, each leaf geometrical shape having a respective 3D shape, the candidate set comprising at least one continuous subset of leaf geometrical shapes, each first candidate leaf geometrical shape of the subset being obtainable by a continuous deformation of second candidate leaf geometrical shape of the subset, the continuous deformation continuously modifying the respective 3D shape of the first candidate leaf geometrical shape into the respective 3D shape of the second candidate leaf geometrical shape, such that each intermediate state of the continuous deformation is itself a candidate leaf geometrical shape of the subset, the intermediate state having a respective 3D shape which is intermediate between the respective 3D shape of the first candidate leaf geometrical shape and the respective 3D shape of the second candidate leaf geometrical shape, and
learn the neural network based on the dataset and on the candidate set such that each leaf geometrical shape of the continuous subset of leaf geometrical shape is available for being potentially inferred during inference by the learnt neural network; and/or
wherein the computer program when executed by the processor causes the processor to be configured to
obtain a discrete geometrical representation of a 3D shape,
apply the neural network to the discrete geometrical representation of a 3D shape, and
infer an editable feature tree representing the 3D shape based on a result of the applying,
wherein the candidate set includes a set product between:
a discrete set of primitive shape types, and
for each primitive shape type, a respective discrete set of one or more parameter domains each of a respective continuous parameter, each parameter domain having respective parameters values of the respective continuous parameter,
each primitive shape type forming, with a respective parameter value for each of the respective discrete set of one or more parameter domains, a respective element of the candidate set, and
wherein for each of one or more primitive shape types, the one or more respective continuous parameters include one or more dimensional parameters, the dimensional parameters being parameters that define a size of characteristic dimensions of a shape of a given primitive shape type.
13. The device of claim 12 , wherein, for each of one or more primitive shape types, the one or more respective continuous parameters include one or more dimensional parameters and/or one or more positioning parameters.
14. The device of claim 12 , wherein the discrete set of primitive shape types include a cuboid type, a sphere type, one or more cylinder types, and/or one or more prism types.
15. The device of claim 12 , wherein the neural network includes recurrent neural network cells, each RNN cell outputting at a respective time step, respective first data for inference of a respective primitive shape type and respective second data for inference of a respective parameter value for each of the respective discrete set of one or more parameter domains.
16. The device of claim 15 , wherein the respective first data include a respective discrete distribution of probabilities each attributed to a respective one of the discrete set of primitive shape types, and/or the respective second data comprise a respective parameter value for each of the respective discrete set of one or more parameter domains.
17. The device of claim 16 , wherein the dataset further includes, for each of one or more discrete geometrical representations, a respective editable feature tree including a tree arrangement of geometrical operations applied to leaf geometrical shapes and representing the 3D shape respective to the discrete geometrical representation, each geometrical shape being formed by a respective primitive shape type with a respective parameter value for each of the respective discrete set of one or more parameter domains, the learning of the neural network comprising a supervised training which includes minimizing a loss L 1 which penalizes, for the time step respective to each leaf geometrical shape of each discrete geometrical representation:
a lowness of the probability of the respective first data attributed to the respective primitive shape type of the leaf geometrical shape, and/or
a disparity between the one or more respective parameter values of the leaf geometrical shape and the one or more respective parameter values of the respective second data.
18. The device of claim 17 , wherein the loss L 1 penalizes:
the lowness with a term of the type
-
∑
n
=
1
N
1
T
^
n
∑
t
=
1
T
^
n
log
[
]
;
and/or
the disparity with a term of the type
∑
n
=
1
N
1
T
^
n
∑
t
=
1
T
^
n
incr
[
d
(
,
)
]
;
where:
n designates said discrete geometrical representation,
t designates said time step,
designates said respective primitive shape type,
designates the probability of the respective first data attributed to ,
designates said one or more respective parameter values of said leaf geometrical shape,
designates said one or more respective parameter values of the respective second data,
d(⋅, ⋅) is a distance function, and
incr[⋅] is an increasing and/or positive function.
19. The device of claim 12 ,
wherein the computer program when executed by the processor causes the processor to be configured to
obtain a dataset including discrete geometrical representations each of a respective 3D shape,
obtain a candidate set of leaf geometrical shapes, the candidate set comprising at least one continuous subset of leaf geometrical shapes,
learn the neural network based on the dataset and on the candidate set,
obtain a discrete geometrical representation of a 3D shape,
apply the neural network to the discrete geometrical representation of a 3D shape, and
infer an editable feature tree representing the 3D shape based on a result of the applying, and optionally refining the inferred editable feature tree.
20. A computer-implemented method for learning a neural network configured for inference, from a discrete geometrical representation of a 3D shape, of an editable feature tree representing the 3D shape, the editable feature tree comprising a tree arrangement of geometrical operations applied to leaf geometrical shapes, the method comprising:
obtaining a dataset including discrete geometrical representations each of a respective 3D shape;
obtaining a candidate set of leaf geometrical shapes, each leaf geometrical shape having a respective 3D shape, the candidate set comprising at least one continuous subset of leaf geometrical shapes, the candidate set being based on a discrete set of primitive shape types and on a respective discrete set of one or more parameter domains for each primitive shape type, each of the one or more parameter domains being of a respective continuous parameter and having respective parameters values of the respective continuous parameter, each respective leaf geometrical shape of the candidate set being obtained by applying a respective parameter value for each of the respective discrete set of one or more parameter domains to a respective primitive shape type, each first candidate leaf geometrical shape of the subset being obtainable by a continuous deformation of second candidate leaf geometrical shape of the subset, the continuous deformation changing continuously the respective parameter value for each of the respective discrete set of one or more parameter domains to apply to the respective primitive shape type of the second candidate leaf geometrical shape, thereby continuously modifying the respective 3D shape of the first candidate leaf geometrical shape into the respective 3D shape of the second candidate leaf geometrical shape, such that each intermediate state of the continuous deformation is itself a candidate leaf geometrical shape of the subset, the intermediate state having a respective 3D shape which is intermediate between the respective 3D shape of the first candidate leaf geometrical shape and the respective 3D shape of the second candidate leaf geometrical shape; and
learning the neural network based on the dataset and on the candidate set such that each leaf geometrical shape of the continuous subset of leaf geometrical shapes is available for being potentially inferred during inference by the learnt neural network,
wherein the candidate set includes a set product between:
a discrete set of primitive shape types, and
for each primitive shape type, a respective discrete set of one or more parameter domains each of a respective continuous parameter, each parameter domain having respective parameters values of the respective continuous parameter,
each primitive shape type forming, with a respective parameter value for each of the respective discrete set of one or more parameter domains, a respective element of the candidate set, and
wherein for each of one or more primitive shape types, the one or more respective continuous parameters include one or more dimensional parameters, the dimensional parameters being parameters that define a size of characteristic dimensions of a shape of a given primitive shape type.Cited by (0)
No later patents cite this yet.
References (0)
No backward citations on record.